Abstract

In many applications it is necessary to perform quantitative scene classification and information extraction on imagery following compression/decompression. An example is the segmentation of multispectral imagery after downlinking. Traditionally, the compression and classification algorithms are designed independently using separate performance metrics, then grafted together. This approach has the advantage of modularity. However, no regard is given to the interaction between the two algorithms which may be significant. An alternative approach is to jointly design both the compression and classification algorithms to maximize their combined performance. This latter formulation has two advantages. The first is that the joint design uses the most pertinent metric of performance—that of the classifier (probability of correct classification). The second is that the interactions between the algorithms are accounted for in the design.

© 1993 Optical Society of America